search

Embracing / GFPose

Official implementation for the CVPR2023 paper: "GFPose: Learning 3D Human Pose Prior with Gradient Fields"
Apache License 2.0
123 stars 5 forks source link

Multi Person #6

Closed gpastal24 closed 5 months ago

gpastal24 commented 5 months ago

Could this method be extended for multi person 3D HPE ?

Embracing commented 5 months ago

Sure. A naive extension would treat multi-person as multiple single persons and solve them using GFPose repeatedly. But better designs should model interactions between multiple people and train GFPose on a multi-person dataset.

gpastal24 commented 5 months ago

Hello thanks for your reply, I was mostly thinking about whether it would be possible to train a variant that outputs lets say N 3D poses directly from N 2D poses or better yet directly from a heatmap representation.

I suppose it is doable.

Embracing commented 5 months ago

Yeah. I agree with you. It is doable and might be worth a try.